Review Reports - A Fast and Efficient Semi-Unsupervised Segmentation and Feature-Extraction Methodology for Artificial Intelligence and Radiomics Applications: A Preliminary Study Applied to Glioblastoma

Round 1

Reviewer 1 Report

An interesting work. The authors are asked to make the following things clearer than now.

1.Explain 'Remove small artifacts' in p.5 with related data and figures showing how it works . Describe its process pictorially.

2. Explain 'Contour identification/filtering and binarization' with related data and figures showing how they work.

3. Two methods have the same error for Patient identifier 3. The compared method outperforms the proposed one for patients identifiers 7 and 29. Why?

Author Response

Reviewer 1

An interesting work. The authors are asked to make the following things clearer than now.

Thank you very much for your positive judgement on our work, “An interesting work“ we really appreciated. In the following we tried to answer to your constructive concerns:

1.Explain 'Remove small artifacts' in p.5 with related data and figures showing how it works . Describe its process pictorially.

[1.2] We provided to add a picture aimed at describe the remove small artifacts and a brief explanation of how it works. Thank you for your valuable suggestion.

Explain 'Contour identification/filtering and binarization' with related data and figures showing how they work.

[1.2] Concerning the filtering a simple gaussian filter with moving window, whose dimension in pixels can be set by the user, was used to reduce the background noise. Then a simple binarization based on adaptive threshold proposed by Liu et al. [1]. No contour identification filter has been considered. After the binarization a simple area growing algorithm is applied. We provided to add some further explanation, thank you. [1] Jihong Liu; Chengyuan Wang An Algorithm for Image Binarization Based on Adaptive Threshold. In Proceedings of the 2009 Chinese Control and Decision Conference; IEEE: Guilin, China, June 2009; pp. 3958–3962.

Two methods have the same error for Patient identifier 3. The compared method outperforms the proposed one for patients identifiers 7 and 29. Why?

[1.3] Yes it is true for patient 7 the reference segmentation tool outperform the new one of about two points of percentage while for patient 29 the obtained results are quite comparable, the reference segmentation method only slightly outperform the new one. Probably this is due to the operator. However, it is worth noticing that despite the slight lower error the new method was able to segment the pathology and that the operator time with the new tool was very limited with respect to the reference segmentation tool.

Reviewer 2 Report

This paper proposes a method for segmentation and feature expraction of tumors. The authors claim that this mehtod is suitable for AI approaches such as SVM classifiers, but have not provided supporting results. The authors also claim that the features extracted through segmentation can be used to train different AI algorithms, but without providing supporting results. Therefore, this paper is incomplete and not convincing. This reviewer would like to know what these AI algorithms are, how they are trained, and what kind improvement can be achieved after training.

Other questions are as follows.

Can "Azienda provinciale per i servizi sanitari" be put into English?

[2.1] We apologize but there is no correspondence in English for this institution. In all official documents we always used the Italian name.

The paper compares the proposed method with only Slicer. What about other well-known methods mentioned in Abstract?

[2.2] Slicer 3D is the most used method adopted by neuroradiologist for the segmentations it is well diffused because it is under GNU license. Other well known methodologies are not free

English writing is not fluent and with many grammatical problems.

[2.3] We tried to solve typos and the grammatical problems you kindly indicated us thank you.

Reviewer 3 Report

Reviewer Comments (Electronics, MDPI)

This article addresses the, A fast and efficient semi-unsupervised segmentation and features extraction methodology for artificial intelligence and radiomics applications: a preliminary study applied to glioblastoma. The proposed paper shows, a simple and effective approach for the semi-unsupervised segmentation and features extraction of brain tumors characterized by high tissue heterogeneity.

1. Remove the dot from figure 7 and 10.

2. Remove the dot from title.

3. Paper looks similar with the previous published paper

[Donelli, M., Espa, G. and Feraco, P., 2022. A Semi-Unsupervised Segmentation Methodology Based on Texture Recognition for Radiomics: A Preliminary Study on Brain Tumours. Electronics, 11(10), p.1573.]

4. Signify the novelty of the proposed scheme.

5. In the introduction section clarify the section details.

6. Figure 9 and 15 description are missing in result section.

7. Describe the novel approaches of the scheme in the comment section.

8. Additionally, the presentation of the manuscript should be reformulated and deeply re-analyzed to present the contribution in a highlighted way, in addition, the conclusions drawn do not examine performance with sufficient analytical detail.

Author Response

Reviewer 3

Thank you for your constructive suggestions aimed at improve the quality of our work. In the following we tried to answer to your constructive concerns:

Remove the dot from figure 7 and 10.

[3.1] Extra dots from figures 7 and 10 removed. Thank you.

Remove the dot from title.

[3.2] Extra dot removed from the title. Thank you.

Paper looks similar with the previous published paper

[3.3] The previous published paper deal only with very simple brain tumors, characterized with tissues homogeneity, namely gliomas. If applied to heterogeneous pathologies such as glioblastomas, which are characterized with different tissues such as necrotic tissues, edema, blood, malignant tissue and vessels the previous published methodology completely fails and it is not able to correctly identify the malignant tissue. The segmentation must be done by hand and it requires an expert neuroradiologist and a lots of time. The methodology proposed in this new work is able to manage heterogeneous structures and it is particularly suitable for glioblastomas pathologies, providing a fast and accurate method for the segmentation of such kind of pathologies. Also an inexpert operator or a student can be trained to use the tool in a very limited amount of time. We provided to emphasize this important aspect in the introduction section. Thank you.

Signify the novelty of the proposed scheme.

[3.4] The main novelty of the new approach is that is able to manage complex and heterogeneous tumors while the previous published approach can only deal with simple pathologies characterized by an homogeneous structures. The new approach is able to manage with an high degree of efficacy homogeneous tumors, such as glioma, as well as very heterogeneous tumors such as glioblastoma. Moreover the new friendly graphical interface permits to simplify the raining phase of inexpert operators. We provided to better specify this important aspect in the body of the manuscript.

In the introduction section clarify the section details.

[3.5] We provided to add the section details. Thank you for your suggestion aimed at improve the readability of the manuscript.

Figure 9 and 15 description are missing in result section.

[3.6] We provided to insert a description for figures 9 and 15 in the result section.

Describe the novel approaches of the scheme in the comment section.

[3.7] Thank you for your constructive suggestion. We provided to better specify the novel approach schema in the comment section, in particular we emphasized the simplified structure of the applied image processing tools and the friendly and direct graphical interface that permits to strongly reduce the operator time burden.

Additionally, the presentation of the manuscript should be reformulated and deeply re-analyzed to present the contribution in a highlighted way, in addition, the conclusions drawn do not examine performance with sufficient analytical detail.

[3.8] Following your constructive suggestion we provided to better highlight the new contribution provided by the proposed methodology in the introduction section and we tried to better examine the obtained performances in the conclusion section. Thank you.

Reviewer 4 Report

This work presented a simple and effective approach for the semi-unsupervised segmentation and features extraction of brain tumors characterized by high tissue heterogeneity. The method is particularly suitable for AI approaches such as SVM classifiers or radiomics applications. Besides, this work also aims to provide a large database of GBMs pathologies for training of neuroradiologist students to learn more about this aggressive disease and provide them software tools aimed to formulate fast and correct diagnosis. The research is meaningful with the article well-structured. However, I have some comments and suggestions for the authors to consider and improve the current manuscript.

1. Could the author envision the potential of this technique in practical applications, as the practical applications requires the trade-off between the high accuracy and applicability of equipment.

2. Will this technique contribute to the earlier diagnosis of tumor?

3. Is this technique specifically aimed to brain tumors? How about tumors in other sites of the body, even in deep tissues?

Author Response

Reviewer 4

Thank you very much for your positive judgement on our work, “The research is meaningful with the article well-structured“ we really appreciated. In the following we tried to improve the quality of our work following your constructive suggestions:

Could the author envision the potential of this technique in practical applications, as the practical applications requires the trade-off between the high accuracy and applicability of equipment.

[4.1] The methodology has been assessed to real data, it operates directly on the DICOM data provided by RM devices, all considered medical data concerns real patients. Expert neuroradiologists as well as medicine students have been trained to use the tool and used it to provide diagnosis that have been also checked by expert clinicians. The conclusion were that the new tool requires a very reduced train, it provided an high accuracy that in most cases outperform the state of the art segmentation tools such as slicer 3D. These are the conclusion of experts in this field. Thanks for letting us to better explain this important aspect.

Will this technique contribute to the earlier diagnosis of tumor?

[4.2] Yes, the goal of the proposed techniques is to provide a fast and accurate segmentation tool for neuroradiologists and make them able to formulate fast, correct and accurate diagnosis. The possibility to analyze a high amount of data with high accuracy, for sure the tool contributes to the earlier diagnosis of tumors, because more patients can be analyzed at the same time. We provided to better clarify this important aspect in the introduction section. Thank you.

Is this technique specifically aimed to brain tumors? How about tumors in other sites of the body, even in deep tissues?

[4.3] Thank you for your suggestion that permit us to better clarify the potentialities of the proposed tool. In this work the proposed tool has been assessed only for brain tumors characterized by high heterogeneity, such as glioblastoma. We used an experimental dataset belonging to the Radiomic Laboratory of the University of Trento and only related to brain pathologies, a previous work was devoted to more simple homogeneous brain pathologies, such as glioma tumors. However, the proposed methodology can be extended/customized to other pathologies located in other sites of the body or in deep tissues. In the future we are plan to use this tool also for other kind of tumors and we are organizing with other different medical staffs to acquire other medical datasets. We provided to add a sentence in the conclusion section to better clarify this important aspect.

Round 2

Reviewer 2 Report

1. Line 2, "Brain surgery"?

2. Line 9, "Artificial intelligence"?

3. Line 18, "other well-known methods"? The paper only compared with one other method. This is the problem I mentioned in my first round comments.

4. Line 36, "provide"?

5. Line 36 & 39, "informations"?

6. Line 39, "areis limited"?

7. Line 54, "methodologies is"?

8. Line 66, "This happen"?

9. Line 69, "a lots of time"?

By reading only 2 pages of the revised manuscript, I found above 9 English writing problems. Obviously, my suggestions made in the first round comments have not been carefully addressed.

Author Response

Dear reviewer we provided to revised the manuscript trying to remove the typos and the minor grammar errors in the body of the manuscript. Thank you very much for indicating us the mistakes that we provided to correct. Thank you.

Concerning the following suggestion "3. Line 18, "other well-known methods"? The paper only compared with one other method. This is the problem I mentioned in my first round comments." we provide to indicate in the abstract that the comparison has been done only with one segmentation method as suggested. Thank you.

Reviewer 3 Report

N/A

Author Response

Thank you very much for your positive judgement on our work and for your constructive suggestions.